Machine Learning You Can Maintain: Observable, Editable, and Modular Models

Talk Abstract Modern machine learning remains dependent on a paradigm in which considerable up-front effort produces models that are monolithic, opaque, and costly to change once trained. In this talk, I will discuss three active lines of research from the AI Division at the Software Engineering Institute that aim to make ML models easier to maintain. First, I will present analytic methods that propagate uncertainty from a neural network’s inputs through to its outputs without sampling. These methods enable applications ranging from efficient online Bayesian updates to assessing how a model responds to rare events. Second, I will describe a machine unlearning method for mixture-of-experts models that isolates the influence of subsets of training data within dedicated experts, so that specific training examples can be removed without retraining from scratch. Finally, I will present techniques for modularizing neural networks that perform end-to-end autonomous control, decomposing them into components for subtasks such as perception and navigation. This approach retains much of the performance of end-to-end learning while allowing individual components to be inspected, tested, and replaced. Together, these efforts point toward ML systems that can be monitored, corrected, and evolved after deployment. Biography Eric Heim is the Chief Scientist of the AI Division at Carnegie Mellon University’s Software Engineering Institute (SEI). The AI Division comprises approximately 50 scientists and engineers whose research advances AI engineering as a discipline. As chief scientist, he provides technical leadership for a portfolio of fundamental and applied research spanning robust and reliable AI, embedded and high-performance AI computing, autonomous systems, AI test and evaluation, agentic AI, and computer vision. He is also active in the AI for humanitarian assistance and disaster response community and has served on the program committee of the HADR+AI workshop at NeurIPS